Understanding Anti-Vaccination Attitudes in Social Media - Microsoft
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Proceedings of the Tenth International AAAI Conference on Web and Social Media (ICWSM 2016) Understanding Anti-Vaccination Attitudes in Social Media Tanushree Mitra1,2 Scott Counts2 James W. Pennebaker2,3 1 2 3 Georgia Institute of Technology Microsoft Research University of Texas at Austin tmitra3@gatech.edu counts@microsoft.com pennebaker@mail.utexas.edu Abstract persistent vaccine criticism movement has spread rapidly The anti-vaccination movement threatens public health by through social media, a channel often used to disseminate reducing the likelihood of disease eradication. With social medical information without verification by the expert media’s purported role in disseminating anti-vaccine infor- medical co mmunity (Keelan et al. 2010). mation, it is imperative to understand the drivers of attitudes Given the increasing reliance on online media for accu- among participants involved in the vaccination debate on a communication channel critical to the movement: Twitter. rate health information and the general growth of social Using four years of longitudinal data capturing vaccine dis- media sites, the attitudes of anti-vaccination advocates risk cussions on Twitter, we identify users who persistently hold becoming a global phenomenon that could impact immun- pro and anti attitudes, and those who newly adopt anti atti- ization behavior at significant scale (Kata 2010). In fact a tudes towards vaccination. After gathering each user’s entire controlled study showed that parents opting to exempt Twitter timeline, totaling to over 3 million tweets, we ex- plore differences in the individual narratives across the user children from vaccination are more likely to have received cohorts. We find that those with long-term anti-vaccination the information online compared to those vaccinating their attitudes manifest conspiratorial thinking, mistrust in gov- kids (Salmon et al. 2005). These parents benefit from “herd ernment, and are resolute and in-group focused in language. immunity” in which eradication is achieved by immunizing New adoptees appear to be predisposed to form anti- a critical proportion of the population. However, as inter- vaccination attitudes via similar government distrust and general paranoia, but are more social and less certain than net-fueled misbeliefs drive people to opt out of vaccina- their long-term counterparts. We discuss how this apparent tion, herd immunity is weakened, increasing the chances of predisposition can interact with social media-fueled events a disease outbreak. Thus it is important to understand the to bring newcomers into the anti-vaccination movement. underlying characteristics of individuals with anti- Given the strong base of conspiratorial thinking underlying vaccination attitudes. What drives people to develop and anti-vaccination attitudes, we conclude by highlighting the need for alternatives to traditional methods of using authori- perpetuate the anti-vaccination movement? tative sources such as the government when correcting mis- In this paper we explore this question by examining in- leading vaccination claims. dividuals’ overt expressions towards vaccination in a social media platform extensively used for vaccine discussions: Twitter. By using four years of longitudinal data capturing Introduction vaccination discussions on Twitter, we identify three sets Measles, a highly contagious respiratory disease responsi- of key individuals: users who are persistently pro vaccine, ble for an estimated 122,000 deaths worldwide each year, those who are persistently anti vaccine and users who new- was officially eradicated from the United States in 2000. ly join the anti-vaccination cohort following an event sym- Yet the disease appears to be rebounding. According to the bolic to the vaccine controversy. Long-term anti- CDC, in 2014 the number of measles cases had reached a vaccination advocates play an important role in preventing 20-year high1(CDC 2015). Sadly, many of these cases eradication because they sustain weakness in herd immuni- could have been prevented, as 90% of measles cases in ty, and thus it is crucial to understand them and their moti- 2014 were in people who were not vaccinated or whose vations. Examining new anti-vaccination proponents al- vaccination status was unknown. One reason for this re- lows us to understand the type of person that would adopt bound is that concerns about vaccine side effects have tak- such a stance despite strong recommendations to the con- en precedence over the dangers of potentially deadly vac- trary from authoritative organizations like the CDC. After cine-preventable diseases and a vaccination culture pro- fetching each cohort’s entire timeline of tweets, totaling to moting anti-vaccination has emerged (Kata 2010). This more than 3 million tweets, we compare and contrast their linguistic styles, topics of interest, social characteristics Copyright © 2016, Association for the Advancement of Artificial Intelli- and underlying cognitive dimensions, all with an eye to gence (www.aaai.org). All rights reserved. 269
Pro vaccination Tweets Anti Vaccination Tweets Measles Outbreak in U.K. Linked to Poor MMR Vaccine Uptake Would rather have measles in my kid than lifetime seizures and autism, After Autism Scare [link] VIT A for a week, big deal #CDCwhistleblower #Vaccines are not a scientific controversy. They work. The vaccines have increased autism, peanut allergies & childhood cancers. Table 1: Example tweets labeled as pro and anti-vaccination by three master Turkers. uncovering the drivers of such extreme attitudes against a this study, we use the evaluative definition of attitude from social good. the seminal work of Eagly and Chaiken (1993) – “attitude We find that people holding persistent anti-vaccination is expressed by evaluating a particular entity with some attitudes use more direct language and have higher expres- degree of favor or disfavor”. By examining the social me- sions of anger compared to their pro counterparts. They dia posts from users we determine whether the expressed also show general conspiracy thinking and mistrust in the attitude towards vaccination is for or against it. government, suggesting characteristics of paranoia. Despite a growing body of work on attitude, a major Adopters of anti-vaccine attitudes show similar conspirato- concern in its study has been the problem of accurate rial ideation and suspicion toward government even before measurement. Attitude studies based on questionnaires and they start expressing anti-vaccine attitudes. This suggests self-reports are highly context dependent and results can that the new adoptees are already predisposed to form anti- vary with changes in question wording, format or order vaccine attitudes. Moreover, long-term anti-vaccine advo- (Schuman and Presser 1981). Participants can also conform cates exhibit higher sense of group solidarity. Individuals to the demands of the questionnaire by creating superficial in such close-knit groups typically end up adhering to their expressions of attitude (Abelson 1988). Even newer meth- extreme positions, often fueling beliefs in false conspira- ods of implicit measures of attitude (Dovidio and Fazio cies and are particularly resistant to correction (Sunstein 1992) have shown sensitivity towards context, raising and Vermeule 2009). These findings suggest that health questions about its effectiveness over self-reported officials attempting to simply correct conspiracy fuelled measures. However, the rapid growth of text-based social false claims might be counterproductive. Thus newer media has opened new opportunities to study attitudes un- methods to counter the harmful consequences of anti- obtrusively, as they naturally unfold in large populations vaccination beliefs are needed. and over long time periods. For example, population atti- Broadly we hope this work contributes to two bodies of tudes extracted from tweet sentiments has been shown to research: computer mediated communication (CMC) re- correlate with traditional polling data (O’Connor et al. search on contentious topics (Kata 2010; Keelan et al. 2010). Machine learning techniques on textual data have 2010; Salmon et al. 2005), and psychology research on accurately predicted sentence level attitudes in online dis- maintaining and adopting an attitude (Kristiansen and cussions (Hassan, Qazvinian, and Radev 2010). Drawing Zanna 1988; Rokeach 1968; Savion 2012). With respect to on the success of studying attitudes from online textual CMC research, examining the naturally occurring self- data, we built a classifier to determine positive and nega- expressions of these cohorts highlight the linguistic style, tive attitudes towards vaccination. social media characteristics, and topics of interest driving people with health behavior beliefs that are antithetical to Analyzing text to infer individual characteristics the health of society as a whole. With regards to attitude Attitudes and language are intimately related (Eiser 1975). psychology, while there are numerous lab-based studies For decades social scientists have demonstrated that indi- listing various factors driving resistant attitudes, our study viduals often adopt language consistent with their attitudes is one of the first to provide large-scale empirical evidence (Eiser and Ross 1977). Hence a growing number of studies of factors behind resistant attitudes towards vaccination. have used content-analytic approaches to assess individual differences and personality characteristics. A popular ap- Related Work proach is simply to count and categorize the words that people speak. A validated tool for such an approach is We provide a brief overview of two broad areas of research LIWC, Linguistic Inquiry and Word Count (Tausczik and relevant to our study: attitude measurement and text analyt- Pennebaker 2010). Researchers have used LIWC for a ic approaches to study user traits. number of psychological measurement tasks, such as deci- phering psychological states of presidential candidates Measuring Attitude from their spontaneous speech samples (Slatcher et al. The concept of attitude has long been central to social psy- 2007), identifying true and false stories by analyzing lin- chology research. However, the definition of attitude has guistic styles (Newman et al. 2003), examining differences changed over the years resulting in no single universally between depressed and non-depressed individuals (Rude, accepted definition (Schwarz 2007). For the purposes of 270
PHASE 1 Stance Classifier vaccine phrases Tweets with Turker labeling labeled Compute features Twitter Firehose Pro/Anti labels Tweets with pro or 49K tweets tweets features SVM probability score > 0.9 Yes from Filter vaccine tweets unlabeled + anti stance Random sample tweets probability score 32K users If pro tweet count > 70% active-pro of all vaccine tweets 373 users PHASE 2 vaccine phrases Check vaccination tweet time Get each user’s Stance pro/anti 653 users vaccine tweets Classifier prob > 0.9 If anti tweet count > 70% active-anti 32K users If vaccine tweets in all 4 years Get twitter timelines Filter vaccine tweets of all vaccine tweets 70 users If vaccine tweets post Aug’14 978 users .... .... .... joining-anti .... 223 users Figure 1: Steps to identify user cohorts. Ellipses (...) shown at the bottom panel of phase 2 correspond to an equivalent module from the top panel. Gortner, and Pennebaker 2004), and contrasting pro- spanning four calendar years – January 1, 2012 to June 30, anorexic and pro-recovery people (De Choudhury 2015). 2015, totaling to 315,240 tweets generated by 144,817 Another popular text analytic technique among social unique users. Our next task was to build a classifier to psychologists is the Meaning Extraction Method (MEM) identify pro and anti stance of the collected vaccine tweets. (Chung and Pennebaker 2008). MEM can reliably infer the Classify Vaccination Stance – Stance Classifier dimensions along which people think about themselves or Our classifier was based on a two-step labeling process: (1) particular issues. Scientists have used MEM to capture gathering human annotations on a sample, and (2) leverage psychological dimensions of self expressions in personal the labeled data to annotate the remaining collection of narratives (Chung and Pennebaker 2008), measure peo- tweets. For the human annotation task, we randomly sam- ple’s basic values underlying their attitude (Ryan L. Boyd pled a batch of 2000 posts from each year (a total of 8000 et al. 2015) and capture dimensions of self-expressions in posts) and recruited 3 independent Amazon Mechanical social media posts, such as Facebook status updates Turk master workers residing in the United States to identi- (Kramer and Chung 2011). Following in their footsteps, fy whether the post content is for or against vaccination. our work uses MEM to capture differences in the dimen- Workers were presented with definitions of pro and anti sions of thought across groups of individuals with differing vaccination content along with example tweets to train attitudes towards vaccination. We complement this with them in the labeling task. We retained posts where all three LIWC’s linguistic analysis to examine differences in ex- workers agreed on the post being either pro or anti. We pression styles. We return to the details of MEM later. show a sample in Table 1. Based on a qualitative examination of the frequently occurring unigrams, bigrams, trigrams and hashtags, we Data Collection found that trigrams and hashtags were prominent cues of a Our data collection involves two main phases. Figure 1 tweet’s stance towards vaccination. For example, phrases outlines the main steps. like vaccines causing encephalitis, already #vaccinein- jured like, #VaccineInducedAutism are indicative of the Phase1: Collecting Pro & Anti-Vaccination Tweets tweet being against vaccination, while shut-up #antivaxx- To fetch vaccine specific posts, we first did a manual ex- ers #science, push mmr vaccine, still deny vaccinations amination of 1000 Twitter posts and 50 news reports on signal pro-vaccination stance. Using trigrams and hashtags vaccination to identify search terms and phrases relevant to as features, we built a supervised learning classifier by the vaccination debate. Based on a snowball sampling ap- training a support vector machine (SVM) under 10-fold proach, we used the initial set of search terms to extract a cross validation. We refer to this as our “stance classifier”. tweet sample from the Twitter Firehose2 stream between We retained only tweets for which the classifier predicted January 1 and 5, 2012. Two researchers then manually stance with high probability (greater than 0.9).The predic- inspected the sample to identify any co-occurring terms tion accuracy of our classifier was 84.7%. The resulting missed and to remove spurious terms that might be return- dataset had 49,354 tweets, classified with attitude stance ing tweets too broad to classify on either side of the issue. and posted by 32,282 unique users. In the next phase, we The final set had 5 phrases: ‘vaccination+autism’, ‘vac- examine this user set to identify our user cohorts. cine+autism’, ‘mmr+vaccination’, ‘measles+autism’ and ‘mmr+vaccine’. Using these phrases we fetched tweets Phase2: Identifying User Cohorts The goal of this phase is to segregate three principle actors 2 – long term advocates of pro and anti vaccination attitude The Twitter Firehose stream is a dataset of all public posts from Twitter made available to us through an agreement with Twitter. and users newly adopting anti-vaccination attitude. Our 271
users with each group contributing about 2.12M, 0.46M and 0.85M tweets respectively. Method What can we learn about people with pro or anti attitudes towards vaccination from their digital footprints left in Figure 2: User cohort statistics Twitter? Recall that people’s natural language expression criterion for selecting long-term advocates is to identify can convey personalities and cognitive processes (Cohn, users who consistently post tweets expressing pro or anti Mehl, and Pennebaker 2004). Hence we adopt a lexical attitudes towards vaccination in all four years (2012 to approach to compare and contrast our cohorts. Lexical ap- 2015). To find users who newly joined the anti-vaccination proaches have been widely used by psychologists to cap- cohort, we needed to fix on a time period before which ture personality traits (Allport and Odbert 1936), cognitive users did not mention about vaccination but actively tweet- processes (Slobin 1996) and more recently to analyze ed anti-vaccination posts after that time. We selected Au- open-ended self-descriptions (Chung and Pennebaker gust 15, 2014 as our cutoff time. This time was marked by 2008). The basic premise of a lexical approach is that the a symbolic event pertaining to the vaccine debate: A false important ways in which individuals differ are represented claim reported that a CDC whistleblower exposed CDC by words. Using lexical approaches we can understand data linking autism among African-American boys follow- how people talk while mentioning the vaccination issue ing MMR vaccination3. The event was widely tweeted un- and what topics they talk about? To investigate the what der the #CDCwhistleblower hashtag and was a topic of aspect we inductively extract meaningful themes from us- active discussion among users interested in the vaccine ers’ natural language expressions using the Meaning Ex- debate. traction Method (Chung and Pennebaker 2008). To exam- ine the how aspect we use the LIWC program for compar- Collecting Historical Posts from Users ing usage rates of function and content words in tweets Recall that 32,282 unique users posted our filtered collec- from individuals. We complement our lexical quantifica- tion of vaccination tweets. We collect their entire Twitter tions with social media specific features such as user’s timeline tracing back from June 30, 2015 (time when our follower to following ratio and tweet volume. data collection started). Using the same set of vaccine re- lated search terms as in phase 1, we search each user’s Meaning Extraction Method (MEM) Twitter timeline to find their history of vaccine posts. Next, we pass their vaccine specific tweets through our MEM is a topic modeling approach to extract dimensions stance classifier, retaining only those for which the classi- along which users express themselves. For example, an fier is able to determine pro or anti stance with high proba- individual with ‘vaccination’ as an important part of her bility (>= 0.9). Finally we classify users as holding pro cognitive self is likely to engage in thoughts related to vac- (anti) attitude if more than 70% of their vaccination tweets cination or even other related health topics (Cantor 1990; are classified as pro (anti). Users who were consistently Markus 1977). These cognitive dimensions will naturally classified as pro or anti across all four years are the long- be reflected in the words that she uses in her everyday so- term advocates. Hereafter we refer to them as active-pro cial media posts. MEM uses a factor analytic approach to and active-anti. Users who were classified as holding an form clusters of co-occurring words (Chung and Penne- anti-vaccine attitude after August 15, 2014 are adoptees of baker 2008). Specifically, it performs a principal compo- the anti-vaccination attitude (the joining-anti cohort). We nent analysis on the matrix of words used by social media intentionally did not create a joining-pro cohort. While post authors to find how they naturally co-occur. These interesting, our focus is on understanding the anti- word co-occurrences identify linguistic dimensions that vaccination users, for whom the active-pro cohort can pro- represent psychologically meaningful themes. Contrary to vide contrast as necessary. other topic modeling approaches, the MEM extracted di- Our cohort selection process takes a very conservative mensions can be examined at the level of an individual, approach, resulting in considerable shrinkage of our cohort allowing us to make meaningful inferences in the context pool. We purposely opt for this strategy to maintain high of a user. Being a word-count based approach, MEM is precision and to have high confidence that the users are highly interpretable. Further, MEM has been shown to indeed holding these extreme attitudes. Our final cohort capture psychological dimensions of self-expressions in comprises 373 active-pro, 70 active-anti, 223 joining-anti personal narratives (Chung and Pennebaker 2008) and in social media posts (Kramer and Chung 2011). 3 http://www.snopes.com/medical/disease/cdcwhistleblower.asp 272
Factors Words Between Groups Within Time war, government, arrest, military, iran, terrorist, Factors Mean Diff p-val Mean Diff p-val 1. Evil Govt. romney, drone, terror, spy, cia, bomb Evil Government A>P Pre A P 0.042 Post > Pre 0.002 voter, republican, candidate, conservative, Family vs. Nuance A>P A 0.012 ns 5. Organic food, organic, product, farm, chemical, healthy, Vaccine & Diseases ns Post > Pre 0.001 Food pesticide, genetically, gmo, fat 6. Family vs. family, interest*, baby, wrong*, kid, toddler, Table 3: Theme comparisons between active-pro (P) and ac- tive-anti (A) groups. ! marks themes where A > P and Nuance mother, argument*, evidence*, belief* ! corresponds to P > A. Statistically significant differences apple, space, nasa, galaxy, smartphone android, are based on an independent sample t-test. ‘ns’ denotes non- 7. Technology significant results. tablet, iphone, rocket, google 8. Vaccine & outbreak, measles, hepatitis, hiv, malaria, vac- Diseases cine, pandemic, infection, death, polio, tb as we would expect long-term advocates to be quite in- volved in discussions of vaccines and diseases. Having Table 2: Themes emerging from the tweets of active-pro and active-anti users. Representative words with the highest factor derived the thematic factors, our next task is to examine loadings are shown per theme. Words with negative factor load- the differences and similarities across these dimensions. ings appear with ‘*’. Following between and within subjects experimental de- sign approach, we perform two sets of comparisons: be- tween groups and within time (see Table 3). Results To start, we compare the two long-term active user co- horts across the 8 factors (Table 3, Between Group). We Understanding Long-Term Advocates find that the active-anti group uses more words referring to We want to identify the characteristics of individuals hold- ‘Evil Government’, ‘Organic Food’ and ‘Family vs. Nu- ing persistent attitudes towards vaccination, our active-pro ance. Given the loadings of the specific words on the and active-anti user cohorts. ‘Family vs. Nuance’ factor, the active-anti cohort appears What topics are relevant to them? To know what topics to be relatively focused on the family aspects of vaccina- the two cohorts talk about, we passed their combined twit- tion but in absolute terms, without reflection on beliefs or ter feeds (2.58M in total) through the Mean Extraction evidence. The active-pro cohort uses ‘Technology’ and helper software (Boyd 2015) which automates the MEM ‘Chronic Health’ topics more, the latter implying a general approach. We identified 8 factors that explain 63.4% of the interest in healthy living. The differences were not statisti- variability in the user data (Table 2)4. Considering the gen- cally significant across the remaining factors – ‘Autism vs. erative nature of language, 63.4% of variance is fairly high Affect’ and ‘Government Vote’. (Chung and Pennebaker 2008). Key factors included the How do the themes differ before and after a user’s first factor labeled ‘Evil Government’, with terms like war, reference to vaccination? This analysis helps us capture government, arrest, military, iran, terrorist. The ‘Chronic any significant difference in a user’s behavior across these Health’ factor comprised terms referring to persistent two important phases. To divide a user’s twitter timeline health conditions and its associated risks and treatment. into the temporal phases, we locate the timestamp of her For the ‘Autism vs. Affect’ factor, words describing emo- first vaccination tweet. Next, we group all tweet content tion loaded negatively, while words referring to autism into two respective buckets – pre posts, comprising of loaded positively, suggesting that this factor is referring to tweets mentioned before user’s first vaccine tweet and post autism in an objective sense, devoid of much emotion. An- content groups tweets after this timestamp (refer Figure 2 other important factor that surfaced is ‘Vaccine & Dis- for count statistics). Recall that our MEM method had al- ease’, with words referring to disease outbreaks (outbreak, ready identified the underlying themes in the entire tweet measles, hepatitis, pandemic, hiv) and vaccination (vac- corpus of the two cohorts. By performing an independent cine, immunization, trial). It is worth noting that the emer- sample t-test across the pre and post groups, we are able to gence of this last factor helps validate the MEM approach, distinguish the themes across time (Table 3, Within Time). We find that users refer to ‘Vaccine & Diseases’, ‘Autism vs. Effect’ and ‘Organic Food’ more in their post time 4 Similar to other topic modeling methods, researchers have some degree tweets compared to pre tweet. This suggests a general of flexibility to determine the number of themes. For MEM, theme inter- pretability is the main determining factor. alertness towards healthy food consumption and the safety 273
Social & Identity Cog thinking & Tentative Do their social media characteristics differ? Finally we Ingroup A > P* Cause P > A* examined the following social media characteristics - 1). Emotion Insight P > A**** Attention status ratio, is the ratio of followers (those who Anxiety P > A* Tentative P > A**** pay attention to the user) to following (those the user pays Anger A > P**** Exclusions P > A**** attention to) (Hutto, Yardi, and Gilbert 2013), (2). Message Linguistic Style Interrogatives P > A**** reach, the ratio of retweets and favorites to the total num- Assent P > A**** Conjunction P > A**** ber of tweets. (3). Amount of directed communication is the Non-fluencies P > A* Certainty A > P* ratio of tweets with “@” mentions to total tweet count. (4). Personal Concerns Informational content index is the ratio of tweets contain- Work P > A* ing a URL. As the distribution of social network based Time Orientation Death A > P**** measures deviates from normality, we compare these Present Focus P > A** Biological States measures across the cohorts using the Wilcoxon signed Engagement Health P > A**** rank test. We find that active-pro had greater attention- Vaccine tweet prop Sexual P > A** status ratio while active-anti users had more message Social Media Characteristics reach. The extent of directed communication and informa- Attention Status P > A* Message Reach A > P* tional content index was not statistically different. Addi- Directed Communication Information content index tionally, to test for differences in the degree of engagement Table 4: Comparing the active-pro (P) and active-anti (A) co- in the vaccine issue, we compute engagement as a propor- horts. Statistically significant LIWC results based on independ- ent sample t-tests are listed. Social media characteristics are tion of tweets related to vaccination in the entire user time- compared using Wilcoxon rank sum test. p-values are shown line. We did not find any statistically significant difference, after BH correction to control familywise error-rate: * p ≤ .05; suggesting that both the long-term advocates were equally ** p ≤ 0.01; ***p ≤ .001; **** p ≤ .0001. vested in forwarding their opinions towards vaccination. of genetically modified food and pesticides following their Understanding Anti-Vaccination Joiners first online expression on vaccines. They also start talking To understand the characteristics of people who join the about autism in a more objective manner, mostly sharing anti-vaccination movement, we compare recent adoptees information about the absence or presence of link between (joining-anti) with those already perpetuating the move- vaccine and autism. ment (active-anti). How do they present themselves? To investigate this, we What topics are relevant to them? To start, we pass the turn to our LIWC results in Table 4. Again, we list only entire Twitter timelines of active-anti and joining-anti us- results with statistically significant differences across co- ers (1.32M posts in total) through the MEM analysis as horts, and call out a few results of note in the text. For in- before. We found that 10 factors, explaining 45.4% of the stance, the active-anti cohort uses significantly more in- variance emerged from our dataset (see Table 5). We label group language, indicating their effort to invoke and main- Factor 1 ‘War News’ because of its reference to terror and tain social connections and group solidarity. Emotional war (terrorist, rebel, Syria). The factor, ‘Secret Govern- differences were noteworthy with the active-pro cohort ment’ contains terms referring to government’s attempt to exhibiting greater anxiety and the active-anti cohort greater conceal information (secret, cia, underground, caught, anger. They also exhibited extreme negative concerns alert). A parallel factor emerging from our analysis is through their higher rates of death related words. On the ‘Government Conspiracy’ with terms alluding to covert contrary, active-pro users focused more on the present, and schemes (bilderberg, imf, conspiracy, politics). Here the had concerns about work and health (higher health and term Bilderberg refers to the Bilderberg group, which has sexual category words). They also showed higher cognitive often been accused of conspiracies (see (“Bilderberg processing through their greater usage of causal and in- Group” 2015)). Another similar theme is the ‘Vaccine sight words, indicated an informal style by using more Fraud’ factor with terms referring to disbelief in vaccina- assents and non-fluencies and displayed lower assuredness tion and government’s attempt to hide the adverse conse- through increased use of tentative, exclusions, interroga- quences of vaccination (coverup, omit, data, cdcwhistle- tives and conjunctions. In contrast, active-anti users were blower). In particular, cdcwhistleblower was a frequently more certain (higher rates of certain words) and more di- used hashtag by people from the anti-vaccination camp to rect in their language (fewer assents and non-fluencies). refer to the popular hoax story linking autism and vaccines They also exhibited a drop in immediacy (less present fo- (“Snopes.com” 2015). cus), suggesting possible emotional distancing, an often Comparing factors across the active-anti and joining- used coping mechanism to deal with acutely upsetting anti user groups shows that joining-anti group refers more events (Holman and Silver 1998). to ‘Vaccine Fraud’, whereas active-anti are more likely to 274
Factors Words Between Groups Within Time terror, war, terrorist, foreign, rebel, regime, Factors Mean Diff p-val Mean Diff p-val 1. War News War News west, turkish, russian, syria, qaeda, humanitarian A>J J J A Pre J Pre 0.002 tax, voter, romney, obama, president, medicare, 4. Govt. Vote Organic Food ns ns democrat, capital, republican, donor, gop, poll Govt. Conspiracy A>J 0.005 ns 5. Vaccine cdcwhistleblower, vaccinate, mmr, fraud, data, War & Terror ns ns Fraud investigate, measles, thomspon, coverup, harm Religious Extremism ns ns 6. Chronic lung, obesity, detect, muscle, clinical, regulation, Table 6: Theme comparisons between the active-anti (A) and Health diabetes, implant, hip, acid, brain, addiction joining-anti (J) groups. Statistically significant differences are 7. Organic chemical, farmer, food, crop, gmo, usda, organic based on an independent sample t-test. !marks themes where A Food environmental, genetically, pesticide, monsanto > J. !correspond to themes where J > A. 8. Government imf, bilderberg, dictatorship, intimidate,eugenic, Conspiracy embassy, laden, conspiracy, politics, infowar Do their social media characteristics differ? We find dispute, catastrophe, osama, felony, nuke, crisis, that active-anti had greater attention-status ratio compared 9. War & Terror helicopter, missile, militia, enforcement, federal to joining-anti. The extent of directed communication, in- 10. Religious hamas, Nazis, german, muslim, union, evil, nige- formational content and message reach was not statistically Extremism ria, accept, warn, refugee, christian, silence different between the two user cohorts. Taken together, these comparisons suggest that those Table 5: Themes emerging from the social media posts of active- joining the anti-vaccination cohort are more social and less anti and joining-anti users. Representative words with the highest factor loadings are shown per theme. definitive – indicators of people who might join a cause or a group. Long-term anti-vaccination supporters who are refer to ‘War News’, ‘Secretive Government’, ‘Govern- concrete and complex in thought had higher attention sta- ment Vote’, ‘Chronic Health’ and ‘Government Conspira- tus ratio – indicators of people who could perpetuate a cy’ (see Table 6). This provides additional evidence that cause. Those joining also posted relatively more content those joining the anti-vaccination camp are specifically about vaccination (higher engagement), suggesting that for concerned about vaccination and a potential vaccine fraud, them this was, at least initially, a specific issue of interest, while the long-term anti-vaccination supporters have a while for long-term anti-vaccination advocates, vaccination broader agenda of government distrust. Comparisons over appears to be one in a number of government conspiracy time show that anti users refer more to ‘Vaccine Fraud’ issues of interest. This general conspiracy thinking and and ‘Chronic Health’ post their first vaccine tweet. cognitive mindset thus appears to provide the perfect net to How do they present themselves? As before, to under- catch people specifically concerned with vaccination. stand how these two groups participate in the vaccine dis- cussion, we turn to comparisons across the LIWC measures (Table 7), calling out a few results of note. The Discussion active-anti users demonstrated higher cognitive complexity Anti-vaccine advocates manifest conspiracy thinking by using complex sentences (based on high preposition Themes emerging from the twitter history of anti-vaccine usage) and exhibited concrete thinking through their in- advocates refer to government conspiracy, deliberate vac- creased use of concrete nouns (articles). In contrast, join- cine frauds, accusations of cover-ups by regulatory bodies ing-anti users signaled lack of definitiveness through high- (‘Secretive Government’) and concerns over terror wars er usage of interrogation, discrepancy, negation, exclu- (‘War News’) (see Table 5). In fact they had significantly sions and conjunctions words. Consistent with their cogni- higher mentions of vaccine fraud and chronic health con- tive concreteness, active-anti users also showed emotional cerns after their first mention of vaccination. This suggests distancing by their lower usage of positive emotion words an important characteristic of people holding unfavorable compared to joining-anti. Also in line with their overly attitude towards vaccination: an inclination towards con- concrete expressions, active-anti group had higher personal spiracy thinking – a way of interpreting the world where concern for money and work. Joining-anti had higher rates conspiracy plays a dominant role (Zonis and Joseph 1994). of leisure words, which is consistent with their higher posi- It is worth noting that these conspiracy related topics sur- tive emotion word usage. They were also more present face as prominent themes amidst the millions of tweets focused and exhibited increased social orientation (higher expressed over multiple years. This suggests that conspira- social process and increased 2nd person pronoun usage). 275
Social & Identity Cog think & Tentative conspiratorial thinking serves as a “cognitive shortcut” that 2ndperson pronoun J > A**** can be used to simplify and explain larger, more complex 3rdperson pronoun J > A** Discrepancy J > A* effects. This sense making function can lead people to con- Imperson Pronoun J > A**** Negation J > A**** spiratorial beliefs despite little evidence to warrant such Social Process J > A**** Exclusions J > A** beliefs (Shermer 2011). This explains the conviction of Affect & Emotional Distance Interrogation J > A**** active-anti cohort towards their anti attitudes over long Positive Emotion J > A* Conjunction J > A* periods of time. A more concerning fact about conspirato- Articles A > J**** Insight J > A**** Prepositions A > J** rial beliefs is their “self-sealing quality” – attempts to re- Time Orientation Personal Concerns ject the theory may backfire and those very attempts may Present Focus J > A**** Leisure J > A**** be characterized as further proof of conspiracy (Sunstein Engagement Money A > J* and Vermeule 2009). Hence conspiratorial beliefs are very Vaccine tweet prop J > A** Work A > J* hard to correct. This suggests that dispelling vaccination Social Media Characteristics myths among long-term anti-vaccine supporters might be Atten Status ratio A > J** Message Reach very difficult, despite the amount of scientific evidence and Directed Communication Information content index rational arguments provided by government officials, sci- Table 7: Comparing the active-anti (A) and joining-anti (J) entific journals or other regulatory bodies (Wolfe 2002). cohorts using LIWC measures. What about emotional appeals to dismiss vaccine myths? We find that complementing their cognitive con- torial ideation is not only driving their anti-vaccination creteness with decreased positive emotion expressions, beliefs but that the very notion of conspiracy has a strong long term anti vaccine advocates can be identified as cate- hold on their way of reasoning about events in the world. gorical thinkers – people whose writing is more focused on “Freedom of information request reveals major government objects, things and categories, marked by higher use of vaccine conspiracy gaia health #autism #aspie” nouns, articles and prepositions (Pennebaker 2013). Such “Chemtrails/Death-Dumps! Secret Govt Operation | categorical thinkers tend to be emotionally distant. In con- USAHM Conspiracy News [link]” trast, active-pro users characterized by higher cognitive “9/11 Blueprint for Truth – The Most Compelling Presenta- processing and an informal expression style signal dynam- tion Proving the 9/11 Conspiracy … [link]” ic thinking (Pennebaker 2013). Thus, appealing to the emo- The above example tweets from our anti users also align tional side of the anti-vaccination movement also is not with research showing people’s consistency with conspira- likely to successfully change their attitudes. torial ideas: someone who believes in one conspiracy theo- Finally, long-term anti-vaccination advocates indicate a ry, tends to believe others as well (Swami et al. 2011). This sense of group solidarity. Social psychologists have shown consistency is an artifact of people’s tendency to maintain that higher ingroup language (like member, family) is in- a coherent system of attitudes so as to strike internal and dicative of group unanimity and a sense of shared identity psychological consistency (Bem 1970). For individuals (Tausczik and Pennebaker 2010). Long-term anti- with anti-vaccine attitudes, a paranoid world of conspiracy vaccination attitude holders’ increased usage of ingroup theories, secret, sinister organizations and manipulative language, greater attention-status ratio and farther message government bodies causing harm are all part of their co- reach are indicative of higher group cohesion, particularly herent system of beliefs. in comparison to their pro-vaccination counterparts. En- hanced group cohesion and message reach are useful prop- Resoluteness towards anti-vaccination stance erties for establishing their beliefs as a movement and to Our results present evidence of the strength and persistence recruiting new members into it. of the underlying attitude conviction of long-term anti- In summary, long-term anti-vaccination supporters are vaccine advocates. They showed relatively higher usage of resolute in their conspiracy worldviews, are categorical concrete nouns, an indication of definitive expression (Ta- thinkers not likely to be appealed to by emotion, and ex- ble 4). They even used more certainty terms compared to hibit high group cohesion. In light of these results, we need their pro counterparts. This aligns with research examining new tactics to counter the damaging consequences of anti- dimensions of attitude strength and resistance towards vaccination beliefs. Drawing from Sunstein’s work on con- change (Pomerantz, Chaiken, and Tordesillas 1995): a spiracy theories, one possible strategy might be to intro- primary dimension of attitude strength is the degree of ex- duce diversity in the closely-knit anti-vaccination groups pressed certainty. In contrast, active-pro advocates used (Sunstein and Vermeule 2009). Recall that anti-vaccination more indirect language. proponents display high ingroup characteristics. Weaken- The emerging MEM themes of conspiratorial worldview ing the well-knit cognitive clusters of extreme theories by also support active-anti cohort’s persistence towards anti- introducing informational and social diversity might reduce vaccine attitudes. Social psychologists have argued that the pool of long-term anti-vaccination advocates. 276
Predisposition among adoptees of anti-vaccine attitudes transient nature of social media sites might surface new Recall that new adoptees of anti-vaccine attitudes reveal search terms not included in our analysis. Hence our results themes indicating conspiracy thinking as well. Most im- are not exhaustive. In this regard, we created a high preci- portantly these themes are captured while tracking the us- sion rather than an all-inclusive but potentially noisy vac- er’s entire twitter history and not just the posts after they cine post dataset. Third, our analysis is based on quantita- start expressing anti-vaccine attitudes. This reveals a sali- tive observations rather than experimentation. Thus we can ent trait of joining-anti users: they see the world through describe “what” we observe, but our “why” explanations the same paranoid lens as the active-anti users. In other are not causal. Future work, both qualitative and experi- words, they are already predisposed to adopt anti-vaccine mentation on why pro or anti vaccine advocates exhibit sentiments. A predisposition is a state of an individual certain attitudes would help deepen the understanding of which when activated by a stimulus makes a person re- this phenomenon. Finally, our results exemplify vaccine spond preferentially to the stimulus (Rokeach 1968). attitudes on just one social media platform: Twitter. We do The CDCwhistleblower controversy is an example of not know how this translates to other sites. We hope that how a social media-driven event can trigger people’s future research can build up on our findings and investigate preexisting disposition to conspiracy thinking and subse- vaccine discussions or more general health information quently provoke their attitudinal expressions that are coun- debates on other social media sites and mainstream media. terproductive to society. Our current analysis cannot draw causal links from the event, but the timeline is highly sug- gestive: (a) people with prior indications of conspiratorial Conclusion thinking start expressing anti-vaccine attitudes towards a Through a case study of the vaccination debate, we demon- topic after an event, where (b) the topic’s anti-attitude is strate how analysis of the natural language expressions and grounded in conspiracy theories and (c) the event itself is social media activities can paint a multi-faceted picture of an alleged conspiracy event. Together this signals the pow- attitudes around a factious topic. Our study of principal er that a single event can exert on attitude formation. In actors in the vaccine debate in a key social media channel this case the attitude is detrimental to society. Additional revealed that long-term anti-vaccination supporters are research to explore and generalize how events impact atti- resolute in their beliefs and they tend toward categorical tudes at population scale and how to control damaging thinking and conspiratorial worldviews. New anti- attitude formation triggered by an event are fruitful areas vaccination adoptees share similar conspiracy thinking and for social media-based social science. hence are predisposed to develop attitudes aligned with the The linguistic analysis results show another characteris- cohort of believers of vaccine myths. These new members tic of new adoptees of anti-vaccine attitudes: their lack of tend to be less assured and more social in nature, but with a assuredness. Perhaps this lack of definitiveness nudges the new and continued focus on health concerns. Our case predisposed minds of joining-anti cohort to latch on to the study suggests that even a single event, CDCWhistleBlow- group of anti-vaccine advocates. We also find that they use er in this case, may be a sufficient trigger for these people more social expressions than their long-term counterparts, to adopt an attitude and join a socially counterproductive again a quality likely to make a person join a group movement. Given these findings, new interventions such as (Gudykunst, Ting-Toomey, and Chua 1988). trying to weaken the long-term government conspiracy base of the movement are needed. Limitations & Future Work Our user cohorts manifest the biases of a population of active social media users engaged in discussing vaccine References and health information online. How representative their Engelmore, R., and Morgan, A. eds. 1986. Blackboard Systems. attitudes and opinions are of the general population is de- Reading, Mass.: Addison-Wesley. batable (Keelan et al. 2010) and this confines the generali- Abelson, Robert P. 1988. “Conviction.” American Psychologist zability of our findings. Studies comparing users and non- 43 (4): 267. users of social media sites have found higher usage among Allport, Gordon W., and Henry S. Odbert. 1936. “Trait-Names: A young adults, no significant gender differences and lower Psycho-Lexical Study.” Psychological Monographs 47 (1): i. percentage of Native American users (Hargittai 2007). Fu- Bem, Daryl J. 1970. “Beliefs, Attitudes, and Human Affairs.,” 34. ture work can examine cohorts who are not heavy social “Bilderberg Group.” 2015. Wikipedia, the Free Encyclopedia. media users but still have strong anti-vaccination views. https://en.wikipedia.org/wiki/Bilderberg_Group. How do such individuals differ from those relying on so- Boyd, R. L. 2015. “MEH: Meaning Extraction Helper (Version cial media to express vaccine opinions? 1.3.02)[Software].” http://meh.ryanb.cc. We were as thorough as possible in choosing the search Boyd, Ryan L., Steven R. Wilson, James W. Pennebaker, Michal phrases used to collect vaccine specific posts. However the Kosinski, David J. Stillwell, and Rada Mihalcea. 2015. “Values in 277
Words: Using Language to Evaluate and Understand Personal O’Connor, Brendan, Ramnath Balasubramanyan, Bryan R. Values.” In Proc. ICWSM. Routledge, and Noah A. Smith. “From Tweets to Polls: Linking Cantor, Nancy. 1990. “From Thought to Behavior: ‘Having’ and Text Sentiment to Public Opinion Time Series.” In. ICWSM. ‘Doing’ in the Study of Personality and Cognition.” American Pennebaker, James W. 2013. The Secret Life of Pronouns: What Psychologist 45 (6): 735. Our Words Say About Us. Reprint edition. Bloomsbury Press. CDC. 2015. “Measles Cases in the United States Reach 20-Year Petty, Richard E., and Jon A. Krosnick. 2014. Attitude Strength: High". http://cdc.gov/media/releases/2014/p0529-measles.html. Antecedents and Consequences. Psychology Press. Chung, Cindy K., and James W. Pennebaker. 2008. “Revealing Pomerantz, Eva M., Shelly Chaiken, and Rosalind S. Tordesillas. Dimensions of Thinking in Open-Ended Self-Descriptions: An 1995. “Attitude Strength and Resistance Processes.” Journal of Automated Meaning Extraction Method for Natural Language.” Personality and Social Psychology 69 (3): 408. Journal of Research in Personality 42 (1): 96–132. Rokeach, Milton. 1968. “Beliefs, Attitudes and Values: A Theory Cohn, Michael A., Matthias R. Mehl, and James W. Pennebaker. of Organization and Change.” 2004. “Linguistic Markers of Psychological Change Surrounding Rude, Stephanie, Eva-Maria Gortner, and James Pennebaker. September 11, 2001.” Psychological Science 15 (10): 687–93. 2004. “Language Use of Depressed and Depression-Vulnerable De Choudhury, Munmun. 2015. “Anorexia on Tumblr: A Charac- College Students.” Cognition & Emotion 18 (8): 1121–33. terization Study.” In Proc. ICDH. Salmon, Daniel A., Lawrence H. Moulton, Saad B. Omer, M. Dovidio, John F., and Russell H. Fazio. 1992. “New Technolo- Patricia deHart, Shannon Stokley, and Neal A. Halsey. 2005. gies for the Direct and Indirect Assessment of Attitudes.” Ques- “Factors Associated with Refusal of Childhood Vaccines among tions about Questions: Inquiries into the Cognitive Bases of Sur- Parents of School-Aged Children: A Case-Control Study.” Ar- veys 21: 204–37. chives of Pediatrics & Adolescent Medicine 159 (5): 470–76. Eagly, Alice H., and Shelly Chaiken. 1993. The Psychology of Savion, Leah. 2012. “Clinging to Discredited Beliefs: The Larger Attitudes. Harcourt Brace Jovanovich College Publishers. Cognitive Story.” Journal of the Scholarship of Teaching and Eiser, J. Richard. 1975. “Attitudes and the Use of Evaluative Learning 9 (1): 81–92. Language: A Two-Way Process*.” Journal for the Theory of Schuman, Howard, and Stanley Presser. 1981. Questions and Social Behaviour 5 (2): 235–48. Answers in Attitude Surveys: Experiments on Question Form, Gudykunst, William B., Stella Ting-Toomey, and Elizabeth Chua. Wording, and Context. Sage. 1988. Culture and Interpersonal Communication. Sage Pubs. Schwarz, Norbert. 2007. “Attitude Construction: Evaluation in Hargittai, Eszter. 2007. “Whose Space? Differences among Users Context.” Social Cognition 25 (5): 638–56. and Non-Users of Social Network Sites.” Journal of Computer- Shermer, Michael. 2011. The Believing Brain: From Ghosts and Mediated Communication 13 (1): 276–97. Gods to Politics and conspiracies—How We Construct Beliefs Hassan, Ahmed, Vahed Qazvinian, and Dragomir Radev. 2010. and Reinforce Them as Truths. Macmillan. “What’s with the Attitude?: Identifying Sentences with Attitude Slatcher, Richard B., Cindy K. Chung, James W. Pennebaker, and in Online Discussions.” In Proc. EMNLP. Lori D. Stone. 2007. “Winning Words: Individual Differences in Holman, E. Alison, and Roxane Cohen Silver. 1998. “Getting‘ Linguistic Style among US Presidential and Vice Presidential Stuck’ in the Past: Temporal Orientation and Coping with Trau- Candidates.” Journal of Research in Personality 41 (1): 63–75. ma.” Journal of Personality and Social Psychology 74 (5): 1146. Slobin, Dan I. 1996. “From thought and language to thinking for Hutto, C. J., Sarita Yardi, and Eric Gilbert. 2013. “A Longitudinal speaking.” In Rethinking Linguistic Relativity, 70–96. 17. Cam- Study of Follow Predictors on Twitter.” In Proc. CHI. bridge University Press. Kata, Anna. 2010. “A Postmodern Pandora’s Box: Anti- Sunstein, Cass R., and Adrian Vermeule. 2009. “Conspiracy The- Vaccination Misinformation on the Internet.” Vaccine 28 (7): ories: Causes and Cures*.” Journal of Political Philosophy 17 1709–16. (2): 202–27. Keelan, Jennifer, Vera Pavri, Ravin Balakrishnan, and Kumanan Swami, Viren, Rebecca Coles, Stefan Stieger, Jakob Pietschnig, Wilson. 2010. “An Analysis of the Human Papilloma Virus Vac- Adrian Furnham, Sherry Rehim, and Martin Voracek. 2011. cine Debate on MySpace Blogs.” Vaccine 28 (6): 1535–40. “Conspiracist Ideation in Britain and Austria: Evidence of a Monological Belief System and Associations between Individual Kramer, Adam DI, and Cindy K. Chung. 2011. “Dimensions of Psychological Differences and Real-World and Fictitious Con- Self-Expression in Facebook Status Updates.” In Proc. ICWSM. spiracy Theories.” British Journal of Psychology 102 (3): 443–63. Kristiansen, Connie M., and Mark P. Zanna. 1988. “Justifying Tausczik, Yla R., and James W. Pennebaker. 2010. “The Psycho- Attitudes by Appealing to Values: A Functional Perspective.” logical Meaning of Words: LIWC and Computerized Text Analy- British Journal of Social Psychology 27 (3): 247–56. sis Methods.” Journal of Language and Social Psychology 29 (1): Markus, Hazel. 1977. “Self-Schemata and Processing Information 24–54. about the Self.” Journal of Personality and Social Psychology 35 Wolfe, Robert M. 2002. “Vaccine Safety Activists on the Inter- (2): 63. net.” Expert Review of Vaccines 1 (3): 249. Newman, Matthew L., James W. Pennebaker, Diane S. Berry, and Zonis, Marvin, and Craig M. Joseph. 1994. “Conspiracy Thinking Jane M. Richards. 2003. “Lying Words: Predicting Deception in the Middle East.” Political Psychology, 443–5. from Linguistic Styles.” Personality and Social Psychology Bul- letin 29 (5): 665–75. 278
You can also read